Summary
Highlights
Gemini 3, Google's new flagship AI model, is now widely available with interesting new features. The presenter has spent several days with it and will cover the big announcements and demonstrate its action. Gemini 3 is flexing its multimodal capabilities, which Google calls the best in the world for multimodal understanding. It's available in Gemini Studio, via API, Vertex, and a new feature called Anti-Gravity.
Gemini 3 tops the LM Arena leaderboard with 1151 points, surpassing Gemini 2.5 Pro by 50 ELO. It scored 37.5% on 'Humanity's Last Exam' without additional tools, which is considered good in this context. It excels in multimodal reasoning with breakthrough scores of 81% and 87.6% on video and multimodal aspects, and 72.1% on simple QA, showing progress in factual accuracy. The task completion success rate for SEMA 2, which uses Gemini 3, is around 65%, a significant improvement from SEMA 1's 31% less than a year ago.
Gemini Agent, based on Project Mariner, can break down requests through deep research, Canvas, Gmail, Calendar, and live web browsing. This generalist agent is hoped to act like a functional Siri. At launch, it will only be available on the Ultra plan. Another feature, Generative Interface, creates interactive presentations from prompt queries, like explaining a Van Gogh gallery or planning a trip to Rome with dynamic sliders and visuals. This is currently an experimental feature.
Google claims Gemini 3 is its most powerful model for coding, topping the WebDev Arena leaderboards. The presenter demonstrated its capabilities by having it build a Missile Command clone and a Mincraft-like voxel world with an SVG of a pelican riding a bicycle. The presenter also had it build 'Hell Maze', a Doom clone, and a guitar tuner that works through the mobile app.
Gemini 3 was tasked with writing the first 10 pages of a supernatural action-horror screenplay about a jewel thief who dies and must steal a heavenly artifact from hell. While it couldn't generate a PDF directly, the script, 'The Devil's Take', was actually not half bad. It moved quickly, introduced the main character's death by page three, and set up a compelling premise with a 'skeleton key of Solomon' as the McGuffin. The character 'Elias' quickly forms a team, including Sarah Vain, a deceased rum runner. While pacing was fast and read like a comic book, the dialogue was better than previous LLMs. Script formatting had some issues when translated, but it's a good tool for testing creative ideas.
Anti-Gravity is Google's agentic development platform for designers. Its agents can open web browsers and check their own work. It includes an 'artifacts' element that keeps receipts of actions taken, providing a checklist of completed tasks. Anti-Gravity is available in public preview at no charge with generous rate limits on Gemini 3 Pro usage.
Gemini 3 Pro is available across all Gemini tiers, a positive step compared to OpenAI. While it may not be Google AGI, it's a significant improvement over 2.5 in just eight months. The video hints at future developments like Nano Banana 2, whispers of V4, and a hopeful release of Genie 3, suggesting more updates are on the horizon.